Neural network

A computing system inspired by the biological neural networks of the human brain, composed of interconnected nodes (neurons) organized in layers that process information, learn patterns from data, and make predictions or decisions.

Source: AIEOG AI Lexicon (Feb 2026), adapted from NIST AI 100-1 and IEEE Standards

A neural network is a type of computer program loosely modeled after how the brain processes information. It consists of layers of mathematical functions — called neurons or nodes — that take in data, transform it through weighted connections, and produce an output. The “learning” happens when the network adjusts these weights based on examples it has seen, gradually improving its ability to recognize patterns, classify data, or make predictions.

Neural networks are the backbone of most modern AI. When you hear about systems that can recognize faces, translate languages, detect fraud, or generate text, neural networks are almost always involved. They range from simple architectures with a few layers to massively complex systems with billions of parameters.

The key distinction from traditional software is that neural networks are not explicitly programmed with rules. Instead, they discover rules from data — which is both their greatest strength and their biggest governance challenge.

Neural networks have become pervasive across financial services, powering systems that directly affect consumers, markets, and institutional risk:

• Credit decisioning. Neural networks can evaluate creditworthiness by identifying complex patterns across hundreds of variables. They often outperform traditional scorecards — but their opacity raises fair lending concerns under ECOA and Regulation B.
• Fraud detection. Transaction monitoring systems increasingly rely on neural networks to identify suspicious patterns in real time. The ability to detect novel fraud schemes is valuable, but the lack of clear decision rationale complicates SAR narrative requirements.
• Trading and market operations. Neural networks drive algorithmic trading strategies, portfolio optimization, and market risk modeling. Their sensitivity to training data and potential for unexpected behavior during market stress events makes robust validation essential.
• Customer service. Chatbots and virtual assistants built on neural networks handle millions of customer interactions. Compliance teams must ensure these systems provide accurate information and appropriate disclosures.
• Anti-money laundering. AML systems use neural networks to score transaction risk and prioritize alerts. The challenge is demonstrating to examiners why certain transactions were flagged and others were not.

• Deep learning — neural networks with many hidden layers
• Machine learning — the broader discipline that includes neural networks
• Transformer — a specific neural network architecture behind modern LLMs
• Foundation models — large neural networks trained on broad data for general-purpose use
• Supervised learning — a training approach commonly used with neural networks